RFM: response-aware feedback mechanism for background based conversation

In general, the existing dialogue systems tend to generate generic responses due to lack of external knowledge. One of the usual solutions is the Background Based Conversations (BBCs), which can help dialogue systems generate more informative and appropriate responses, based on an external knowledge source. Unfortunately, there still exists some difficulties for BBCs when correcting the selected knowledge during response generation, e.g., see GTTP, CaKe. In this paper, we propose a novel architecture called Response-aware Feedback Mechanism (RFM) for BBCs to address this shortcoming. The main advantage is that a Response-aware Feedback Weight Vector is introduced to integrate the background knowledge and responses, so that the knowledge selector could select more accurate knowledge. With the help of this self-correcting mechanism, the selected knowledge is adjusted and corrected dynamically in each decoding time step. As an application, we carry out experiments on the Holl-E and Wizard of Wikipedia datasets, the results indicate that the RFM model has much better performance on automatic and human evaluation, compare to eleven state-of-the-art methods, including RefNet and GLKS.